Research map / Point clouds / Point cloud classification
CurveCloudNet: Processing Point Clouds with 1D Structure
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CurveCloudNet, a point cloud backbone that parameterizes LiDAR/laser-scan point clouds as polylines ("curve clouds") and applies 1D curve operations at high resolution mixed with point operations at low resolution.
Problem
Point-based 3D backbones scale poorly to large scenes, while sparse-voxel backbones discretize small or irregular point sets and lose detail; existing methods ignore the 1D curve-like traversal structure of laser-scanner outputs.
Approach
Link LiDAR/laser points into polylines using beam IDs, timestamps and distance thresholds to form a curve cloud, then apply 1D operations along curves (geodesic farthest point sampling, curve grouping, symmetric 1D convolutions) in a U-Net that uses curve layers at high resolution and point layers at low resolution.
Key results
- CurveCloudNet reaches 72.7 average mIoU across five datasets, the best average in Table 1 (vs 70.6 for SphereFormer).
- On A2D2 it achieves 54.1 mIoU, outperforming point-based PointMLP (47.6) and competitive with SphereFormer (55.1).
- On ShapeNet it reaches 83.1 class-average mIoU, the highest among compared methods, and on KortX test 78.9 mIoU versus 73.7 for PointNext.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CurveCloudNet instead of CurveNet | ShapeNet part segmentation · ShapeNet | mIOU: 82.8 -> 83.1 | Helped |
| use CurveCloudNet instead of PointMLP | KortX part segmentation · KortX | mIOU: 75.4 -> 78.9 | Helped |
| use CurveCloudNet instead of SphereFormer | A2D2 semantic segmentation · A2D2 | mIoU: 55.1 -> 54.1 | Hurt |
| use CurveCloudNet instead of SphereFormer | nuScenes semantic segmentation · nuScenes validation | mIoU: 79.5 -> 78.0 | Hurt |
| use CurveCloudNet instead of SphereFormer | KITTI semantic segmentation · KITTI validation | mIoU: 69.0 -> 69.5 | Helped |
| add curve grouping instead of regular radial grouping (ablation) | CurveCloudNet · A2D2 | mIoU: 53.3 -> 54.1 | Helped |
| add curve FPS instead of regular FPS (ablation) | CurveCloudNet · A2D2 | mIoU: 52.4 -> 54.1 | Helped |
| add 1D curve convolutions (ablation) | CurveCloudNet · A2D2 | mIoU: 52.0 -> 54.1 | Helped |
| add curve operations instead of point-based counterparts (ablation) | CurveCloudNet · A2D2 | mIoU: 52.6 -> 54.1 | Helped |
Surprising findings
- Voxel-based backbones underperform their point-based counterparts on small object point clouds, suggesting discretization hurts when point sets are small.
- CurveNet, a prior curve-based method, reaches only 4.4 mIoU on A2D2 and exceeds 48GB GPU memory on nuScenes, failing to scale.
Limitations
- Designed only for laser-scanned data with explicit curve structure; does not directly apply to uniformly sampled point clouds.
- Some baselines (e.g., CurveNet, PointNet++ on KITTI) could not be trained due to excessive training time, so comparisons are incomplete.
- KortX dataset is small (195 real scans) and only released upon publication.
Experimental setup
Datasets: ShapeNet KortX A2D2 nuScenes SemanticKITTI
Benchmarks: ShapeNet Part Segmentation KortX Part Segmentation A2D2 nuScenes SemanticKITTI
Metrics: mIoU latency (ms) GPU memory (GB) parameters (M)
Baselines: PointNet++ CurveNet PointMLP PointNext DGCNN MinkowskiNet Cylinder3D SphereFormer RandLANet PolarNet
Builds on: PointNet++ PointNext PointMLP DGCNN MinkowskiNet
Closest papers by meaning
- Points-to-3D: Structure-Aware 3D Generation with Point Cloud Priors (CVPR 2026 · similarity 0.88)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.87)
- DANCE: Density-agnostic and Class-aware Network for Point Cloud Completion (AAAI 2026 · similarity 0.87)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.87)
- GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds (CVPR 2024 · similarity 0.87)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.